## Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

**doi:10.1093/imamci/dnn003
**

On robust stability of linear neutral systems with time-varying delays

EMILIA FRIDMAN†

Department of Electrical Engineering-Systems, Tel Aviv University, Tel Aviv 69978, Israel

[Received on 13 June 2007; accepted on 20 January 2008]

The application of the direct Lyapunov method to the stability analysis of neutral systems with time-

varying delays usually encounters a restrictive assumption on the function in the right side of the

differential equation. This function is supposed to satisfy the Lipschitz condition with respect to the

delayed state derivative with a constant less than 1. In the present paper, we extend the input–output

approach to consider the stability of neutral type systems with uncertain time-varying delays and norm-

bounded uncertainties. The assumption on the Lipschitzian constant can then be avoided. Sufﬁcient

stability criteria are derived in the frequency domain and the time domain, where the descriptor

discretized Lyapunov–Krasovskii functional is applied. As a by-product, new necessary conditions for

neutral-delay-independent/retarded-delay-dependent stability criteria are obtained. The method can be

easily extended to L

2

-gain analysis and can be applied to design problems.

Keywords: neutral system; time-varying delay; input–output approach; Lyapunov–Krasovskii method.

1. Introduction

There are two main methods for the stability analysis of linear systems with delay: the direct Lyapunov

method and the input–output approach, based on the small-gain theorem (see, e.g. Gu et al., 2003).

The latter approach has been applied to robust stability analysis of linear ‘retarded-type’ systems with

norm-bounded uncertainties and uncertain time delays in Huang & Zhou (2000), Gu et al. (2003),

Kao & Lincoln (2004) and Fridman & Shaked (2006).

The application of the direct Lyapunov method to neutral type systems with ‘constant neutral de-

lays’ requires a well-known assumption on the stability of the difference equation (Hale & Lunel, 1993;

Kolmanovskii & Myshkis, 1999; Niculescu, 2001). Necessary and sufﬁcient conditions for the stability

of this difference equation are given in Hale & Lunel (1993) in the frequency domain. It was shown in

Fridman (2002) that the descriptor approach to neutral systems with constant delays (see, e.g.

Fridman, 2001) implies the stability of the difference equation and thus avoids a veriﬁcation of the

above assumption. In the case of time-varying neutral delays, the situation becomes more difﬁcult and

the only known assumption for the application of direct Lyapunov method may be rather conservative:

the sum of the norms of the matrices, which multiply the delayed state derivatives, should be less than 1

(El’sgol’ts & Norkin, 1973; Kolmanovskii & Myshkis, 1999). The latter assumption becomes especially

restrictive in the case of an uncertain neutral part or in the case of multiple delays in the derivative of

the state.

In the present paper, a new method is developed for the stability analysis of linear ‘neutral sys-

tems with time-varying delays and norm-bounded uncertainties’ in the neutral part. To the best of the

author’s knowledge, this is the ﬁrst method that avoids the restrictive assumption mentioned above.

†

Email: emilia@eng.tau.ac.il

c The author 2008. Published by Oxford University Press on behalf of the Institute of Mathematics and its Applications. All rights reserved.

2 of 15 E. FRIDMAN

The presented method extends the input–output approach to uncertain neutral systems with time-varying

delays and it leads to sufﬁcient frequency-domain stability conditions. Also some ‘necessary conditions’

for the feasibility of the resulting stability conditions are deduced. To illustrate the efﬁciency of the new

method for neutral systems, the time-domain results are derived via the descriptor discretized Lyapunov

functional method (Fridman, 2006). The latter method combines the discretized Lyapunov functional

method of Gu (1997) with the descriptor model transformation (Fridman, 2001) and it efﬁciently solves

design problems (for the ﬁrst time for the discretized method).

The descriptor discretized Lyapunov functional method is chosen because of the power of the dis-

cretized method (which, differently to the simple Lyapunov functional methods, can analyse systems

which are not stable without delays). We note that in Fridman (2006), the case of constant neutral de-

lays and fast-varying state delays (without any constraints on the delay derivative) was studied. In the

present paper, we extend the method of Fridman (2006) to the case of slowly varying delays in the state

and in the state derivative (where the delay derivative is bounded from above by a constant, which is

less than 1).

Notation: Throughout the paper, the superscript ‘’ stands for matrix transposition, R

n

denotes the

n-dimensional Euclidean space with vector norm· , R

n×m

is the set of all n ×m real matrices and the

notation P >0, for P ∈ R

n×n

, means that P is symmetric and positive deﬁnite. We also denote x

t

(θ) =

x(t +θ)(θ ∈ [−h −µ, 0]). The symmetric elements of the symmetric matrix will be denoted by ∗. L

2

is

the space of square-integrable functions v: [0, ∞) → C

n

with the norm v

L

2

=

¸

∞

0

v(t )

2

dt

¸

1/2

.

A denotes the Euclidean norm of an n×n (real or complex) matrix A, which is equal to the maximum

singular value of A. For a transfer function matrix of a stable system G(s), s ∈ C,

G

∞

= sup

−∞<w<∞

G(iw), i =

√

−1.

σ(B) is the spectral radius of matrix B (i.e. the maximum absolute value of its eigenvalues).

2. Retarded-delay-dependent/neutral-delay-independent stability conditions

2.1 Problem formulation

Consider a linear system

˙ x(t ) = (A

0

+ HE

0

)x(t ) +(A

1

+ HE

1

)x(t −τ(t )) +(F + HE

2

) ˙ x(t − g(t )), (2.1)

where x(t ) ∈ R

n

, A

0

, A

1

, F and E

i

, i = 0, 1, 2, are constant matrices. (t ) is a time-varying uncertain

n ×n matrix that satisﬁes

(t )(t ) I

n

. (2.2)

The uncertain delays τ(t ) and g(t ) are differentiable functions of the form

τ(t ) = h +η(t ), |η(t )| µ h, ˙ τ(t ) d < 1, ˙ g(t ) f < 1, (2.3)

with the known bounds µ, d and f . The neutral delay g(t ) does not have bound other than the derivative

bound.

If τ(t ) = g(t ), then the stability criterion may be applied with f = d. For simplicity, we consider

a single delay g(t ). The presented results may be easily ‘generalized to the case of any ﬁnite number of

delays g(t )’.

ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 3 of 15

The stability analysis of systems with a neutral time-varying delay is a classical problem(El’sgol’ts &

Norkin, 1973). In the robust stability context, where the parameter uncertainties and the variation of the

time delay are taken into account, this problembecomes especially important. In the past, robust stability

of (2.1) with time-varying delay was studied under the restrictive assumption (Lien, 2005)

F +HE

2

< 1, (2.4)

where · is any matrix norm. The latter assumption (which becomes especially restrictive in the case

of multiple neutral delays) allows the direct Lyapunov method to be applied where it is assumed that

the right side of (2.1) (denoted by f (t, x

t

, ˙ x

t

)) satisﬁes the Lipschitz condition with respect to ˙ x

t

with a

constant less than 1 (see Kolmanovskii & Myshkis, 1999, p. 336). Inequality (2.4) guarantees the latter

Lipschitz condition since

f (t, x

t

, ˙ x

t

) − f (t, x

t

,

˙

¯ x

t

) =(F + HE

2

)[ ˙ x(t − g(t )) −

˙

¯ x(t − g(t ))]

(F +HE

2

)˙ x(t − g(t )) −

˙

¯ x(t − g(t )).

In some papers (see, e.g. Park, 2002) from the fact that

˙

V < 0, where V > 0 is a Lyapunov func-

tional, it is directly concluded that the neutral system is asymptotically stable (without any references to

the corresponding Lyapunov theorems). Unfortunately, such a conclusion is not correct.

In the present paper, we develop the input–output approach to the stability analysis of uncertain

neutral type systems with time-varying delays. We ﬁrst derive frequency-domain stability conditions

and then deduce their implications. Further, we ﬁnd time-domain conditions by applying descriptor

discretized Lyapunov functional method. Finally, we illustrate the efﬁciency of the proposed method by

numerical examples. The new method gives tools for solution of different robust control problems for a

wide class of neutral systems, where the condition (2.4) is removed.

2.2 Frequency-domain stability conditions

Representing

x(t −τ(t )) = x(t −h) −

t −h

t −h−η(t )

˙ x(s)ds

and applying the input–output approach (see Gu et al., 2003 and the references therein), we consider the

following forward system:

˙ x(t ) = A

0

x(t ) + A

1

x(t −h) +µA

1

v

1

(t ) + Fv

2

(t ) + Hv

4

(t ),

y

1

(t ) = ˙ x(t ), y

2

(t ) =

1

√

1 − f

˙ x(t ), y

3

(t ) =

1

√

1 −d

x(t ),

y

4

(t ) = E

0

x(t ) + E

1

x(t −h) +µE

1

v

1

(t ) + E

2

v

2

(t ),

(2.5a–d)

with the feedback

v

1

(t ) = −

1

µ

−h

−h−η(t )

y

1

(t +s)ds, v

2

(t ) =

√

1 − f y

2

(t − g(t )),

v

3

(t ) =

√

1 −dy

3

(t −τ(t )), v

4

(t ) = y

4

(t ). (2.6)

4 of 15 E. FRIDMAN

Similarly to He et al. (2004), we add to (2.5a) the left side of the equation

C[x(t −h) +µv

1

(t ) −v

3

(t )] = C

¸

x(t −h) −

−h

−h−η(t )

˙ x(t +s)ds − x(t −τ(t ))

¸

= 0,

where C is an arbitrary n × n matrix (which is equivalent to the parameterized model transformation

Niculescu, 2001). We represent the forward system in the following parameterized form:

˙ x(t ) = A

0

x(t ) +(A

1

+C)x(t −h) +µ(A

1

+C)v

1

(t ) + Fv

2

(t )

−Cv

3

(t ) + Hv

4

(t ),

y

1

(t ) = ˙ x(t ), y

2

(t ) =

1

√

1 − f

˙ x(t ), y

3

(t ) =

1

√

1 −d

x(t ),

y

4

(t ) = E

0

x(t ) + E

1

x(t −h) +µE

1

v

1

(t ) + E

2

v

2

(t ).

(2.7a–d)

Note that C = 0 corresponds to the moderately varying delay τ(t ) with ˙ τ 1 (Fridman & Shaked,

2006), while C = − A

1

corresponds to τ-independent/ ˙ τ-dependent result. The input–output model

(2.7a–d), (2.6) and the results of the present paper are appropriate also to the system (2.1) with delay

h 0 and non-negative η ∈ [0, µ]. In the latter case, C = 0 corresponds to systems with fast-varying

delay τ (i.e. without any constraints on ˙ τ).

In the case of a retarded system with F = 0 and with ˙ τ 1, the input–output model (2.7a–d), (2.6)

has been introduced in Fridman & Shaked (2006), where C, v

i

and y

i

, i = 2, 3, were taken to be zero.

Moreover, y

1

and y

2

correspond to the descriptor method (Fridman, 2001), where ˙ x(t ) appears in the

derivative of the Lyapunov functional.

We assume the following.

(A1) The (parameterized) nominal system

˙ x(t ) = A

0

x(t ) +(A

1

+C)x(t −h) (2.8)

is asymptotically stable.

Let v

= [v

1

· · · v

4

] and y

= [y

1

· · · y

4

]. Assume that y

i

(t ) = 0, ∀ t 0, i =

1, . . . , 4. The following holds:

v

i

L

2

y

i

L

2

, i = 1, . . . , 4. (2.9)

The forward system (2.7a–d) can be written as y = Gv with transfer matrix

G(s) =

¸

s I

n

1

√

1 − f

s I

n

1

√

1 −d

I

n

E

0

+ E

1

e

−hs

¸

×(s I − A

0

−(A

1

+C)e

−hs

)

−1

[µ(A

1

+C) F −C H]

+

¸

¸

0

3n×n

0

3n×n

0

3n×2n

µE

1

E

2

0

n×2n

. (2.10)

ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 5 of 15

By the small-gain theorem (see, e.g. Gu et al., 2003), the system (2.1) is input–output stable (and

thus asymptotically stable, since the nominal system is time invariant) if G

∞

< 1. A stronger result

may be obtained by scaling G.

THEOREM 1 Consider (2.1) with delays given by (2.3), where η(t ) and g(t ) are differentiable functions.

If there exists an n ×n matrix C such that A1 holds and there exist non-singular n ×n matrices X

i

, i =

1, 2, 3, and a scalar r = 0 such that

G

X

∞

< 1, G

X

(s) = diag{X

1

, X

2

, X

3

, r I

n

}G(s) diag{X

−1

1

, X

−1

2

, X

−1

3

, r

−1

I

n

}, (2.11)

then (2.1) is input–output stable.

REMARK 1 In Section 2.4 below sufﬁcient conditions for the feasibility of A1 and (2.11) will be derived

in terms of linear matrix inequalities (LMIs). The free matrices C, X

i

and the scalar r of (2.11) will be

related to the decision variables of the LMIs (see Remark 3 below).

2.3 Implications of the frequency-domain stability conditions

Now, we are in a position to formulate some necessary conditions for the feasibility of (2.11).

PROPOSITION 1 If (2.11) holds, then

(i) the eigenvalues of µ(A

1

+ C) are inside the unit circle (µσ(A

1

+ C) < 1), i.e. the difference

equation

x(t ) −µ(A

1

+C)x(t − g

0

) = 0, (2.12)

with constant delay g

0

is stable;

(ii) the eigenvalues of

1

√

1−f

F are inside the unit circle, i.e. the difference equation

x(t ) −

1

√

1 − f

Fx(t − g) = 0, (2.13)

with constant delay g is stable;

(iii) σ

0

< 1, where

σ

0

= sup

¸

σ

µ(A

1

+C)e

iθ

0

+

1

√

1 − f

F e

iθ

¸

: θ

0

, θ ∈ [0, 2π]

¸

, (2.14)

which is equivalent to the delay-independent stability of the following difference equations with

constant delays g

0

and g (Hale & Lunel, 2003, Theorem 6.1, p. 286):

x(t ) −µ(A

1

+C)x(t − g

0

) −

1

√

1 − f

Fx(t − g) = 0. (2.15)

6 of 15 E. FRIDMAN

Proof. We will prove (iii) only. The proof of (i) and (ii) is similar.

Assume that (2.11) holds. Then,

¸

¸

¸

¸

¸

¸

X

1

X

2

√

1−f

[µ(A

1

+C)X

−1

1

FX

−1

2

]

¸

¸

¸

¸

¸

sup

−∞<w<∞

¸

¸

¸

¸

¸

iw

¸

X

1

X

2

√

1−f

(iwI

n

−A

0

−(A

1

+C)e

−hiw

)

−1

× [µ(A

1

+C)X

−1

1

FX

−1

2

]

¸

¸

¸

¸

¸

G

X

∞

< 1. (2.16)

Let λ and a be an eigenvalue and an eigenvector of

µ(A

1

+C)e

iθ

0

+

F

√

1−f

e

iθ

, i.e.

µ(A

1

+C)e

iθ

0

+

F

√

1 − f

e

iθ

¸

a = λa.

From (2.16), it follows that

|λ|

¸

¸

¸

¸

¸

¸

X

1

a

X

2

a

√

1−f

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

¸

X

1

X

2

√

1−f

λa

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

¸

X

1

X

2

√

1−f

[µ(A

1

+C)X

−1

1

FX

−1

2

]

¸

X

1

e

iθ

0

a

X

2

√

1−f

e

iθ

a

¸

¸

¸

¸

¸

<

¸

¸

¸

¸

¸

¸

X

1

e

iθ

0

a

X

2

√

1−f

e

iθ

a

¸

¸

¸

¸

¸

=

¸

¸

¸

¸

¸

¸

X

1

a

X

2

√

1−f

a

¸

¸

¸

¸

¸

. (2.17)

Hence, |λ| < 1. This completes the proof.

REMARK 2 We note that the existing LMI stability criteria, derived via different direct Lyapunov meth-

ods, may usually be recovered via an input–output approach (similar to Zhang et al., 2001; Gu et al.,

2003). In this case, the LMI conditions (which give sufﬁcient conditions for the frequency-domain con-

dition (2.11) and for A1) are feasible if the difference equation (2.15) is stable, where C is related to the

decision variables of these LMIs (as, e.g. in (2.34) below).

Necessary conditions for delay-dependent stability via different model transformations of linear re-

tarded type systems with ‘small’ delays (where h = 0 and η ∈ [0, µ]) were found in Gu & Niculescu

(2001) and Kharitonov & Melchor-Aguilar (2002), where additional dynamics of the transformed

systems were analysed. A simple condition of Gu & Niculescu (2001) (for the constant delay case)

µσ(A

1

) <1 coincides with (i) for the case of fast-varying delay. However, for h = µ and ˙ τ 1 (i)

guarantees the stability on the double interval τ(t ) ∈ [0, 2µ]. A simple condition of Kharitonov &

Melchor-Aguilar (2002)

¸

m

k=1

µ

k

A

k

< 1 for the system

˙ x(t ) =

m

¸

k=1

A

k

x(t −τ

k

(t )), τ

k

(t ) ∈ [0, µ

k

], ˙ τ

k

< 1,

ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 7 of 15

is more restrictive than (i) in the case of ˙ τ

k

1, where the latter has the form

sup

¸

σ

m

¸

k=1

µ

k

A

k

e

iθ

k

¸

: θ

k

∈ [0, 2π]

¸

< 1

and guarantees the stability on the double intervals τ

k

∈ [0, 2µ

k

] for h

k

= µ

k

.

2.4 Time-domain criterion for robust stability

We will derive sufﬁcient conditions for (2.11) by using the complete Lyapunov–Krasovskii functional

(LKF):

V(x

t

) = x

(t )P

1

x(t ) +2x

(t )

0

−h

Q(ξ)x(t +ξ)dξ +

0

−h

0

−h

x

(t +s)R(s, ξ)dsx(t +ξ)dξ

+

0

−h

x

(t +ξ)S(ξ)x(t +ξ)dξ, P

1

> 0, (2.18)

which for S ≡ 0 corresponds to necessary and sufﬁcient conditions for stability of the nominal system

(2.8). The time-domain results will be derived via the discretized Lyapunov functional method (Gu,

2003), since this powerful method (differently from simple Lyapunov functionals) can be applied to

the systems, which are not stable without delays. The new descriptor discretized Lyapunov method of

Fridman (2006) will be used, which allows to solve the design problems.

Since

∞

0

v

(t )

¯

Xv(t )dt

∞

0

y

(t )

¯

Xy(t )dt,

¯

X = diag{µR

a

, (1 − f )U, (1 −d)S

a

, ρI

n

},

the following condition along with (2.7a–d)

W

=

d

dt

V(x

t

) +(1 − f )y

2

(t )Uy

2

(t ) +µy

1

(t )R

a

y

1

(t )

+(1 −d)y

3

(t )S

a

y

3

(t ) +ρy

4

(t )y

4

(t ) −(1 − f )v

2

(t )Uv

2

(t )

−µv

1

(t )R

a

v

1

(t ) −(1 −d)v

3

(t )S

a

v

3

(t ) −ρv

4

(t )v

4

(t )

< −ε(x(t )

2

+˙ x(t )

2

+v(t )

2

), ε > 0, (2.19)

for some n × n matrices R

a

> 0, U > 0, S

a

and a scalar ρ > 0 guarantees the asymptotic stability of

(2.1) (Gu et al., 2003). Note that (2.19) guarantees A1. Therefore, in the time domain we do not assume

the asymptotic stability of the nominal system.

8 of 15 E. FRIDMAN

Differentiating V(x

t

) along the trajectories of (2.7a–d), we obtain that

˙

V is given by

˙

V(x

t

) = 2˙ x

(t )

¸

P

1

x(t ) +

0

−h

Q(ξ)x(t +ξ)dξ

+2x

(t )

0

−h

Q(ξ) ˙ x(t +ξ)dξ

+2

0

−h

0

−h

˙ x

(t +s)R(s, ξ)ds x(t +ξ)dξ +2

0

−h

˙ x

(t +ξ)S(ξ)x(t +ξ)dξ. (2.20)

Adding to

˙

V(x

t

) the right side of the expression

0 = 2[x

(t )P

2

˙ x

(t )P

3

]

×

¸

A

0

x(t ) − ˙ x(t ) +(A

1

+

¯

C)x(t −h) +µ(A

1

+

¯

C)v

1

(t ) + Fv

2

(t ) −

¯

Cv

3

(t ) + Hv

4

(t )

A

0

x(t ) − ˙ x(t ) +(A

1

+C)x(t −h) +µ(A

1

+C)v

1

(t ) + Fv

2

(t ) −Cv

3

(t ) + Hv

4

(t )

¸

,

(2.21)

where P

2

and P

3

are n ×n matrices, which is equivalent to descriptor model transformation of Fridman

(2001) and integrating by parts in (2.20), we ﬁnd

˙

V(x

t

) = ζ

Ξζ +2˙ x

(t )

0

−h

Q(ξ)x(t +ξ)dξ

−

0

−h

0

−h

x

(t +ξ)

∂

∂ξ

R(ξ, θ) +

∂

∂θ

R(ξ, θ)

¸

x(t +θ)dθ dξ

+2x

(t )

0

−h

[−

˙

Q(ξ) + R(0, ξ)]x(t +ξ)dξ

−2x

(t −h)

0

−h

R(−h, θ)x(t +θ)dθ −

0

−h

x

(t +ξ)

˙

S(ξ)x(t +ξ)dξ

+2[x

(t )P

2

˙ x

(t )P

3

]

¸

µ(A

1

+

¯

C)v

1

(t ) + Fv

2

(t ) −

¯

Cv

3

(t ) + Hv

4

(t )

µ(A

1

+C)v

1

(t ) + Fv

2

(t ) −Cv

3

(t ) + Hv

4

(t )

¸

, (2.22)

where

ζ =

¸

¸

¸

x(t )

˙ x(t )

x(t −h)

, Ξ =

¸

¸

¸

Ψ

¸

P

2

(A

1

+

¯

C)

P

3

(A

1

+C)

−

¸

Q(−h)

0

¸

∗ −S(−h)

,

P =

¸

P

1

0

P

2

P

3

¸

,

Ψ = P

¸

0 0

A

0

−I

¸

+

¸

0 A

0

0 −I

¸

P +

¸

Q(0) + Q

(0) + S(0) 0

0 0

¸

. (2.23a–c)

We apply next the discretization of Gu (1997). Divide the delay interval [−h, 0] into N segments

[θ

p

, θ

p−1

], p = 1, . . . , N, of equal length

¯

h = h/N, where θ

p

= −p

¯

h. This divides the square

[−h, 0] ×[−h, 0] into N ×N small squares [θ

p

, θ

p−1

] ×[θ

q

, θ

q−1

]. Each small square is further divided

into two triangles.

ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 9 of 15

The continuous matrix functions Q(ξ) and S(ξ) are chosen to be linear within each segment and the

continuous matrix function R(ξ, θ) is chosen to be linear within each triangular:

Q(θ

p

+α

¯

h) = (1 −α)Q

p

+αQ

p−1

,

S(θ

p

+α

¯

h) = (1 −α)S

p

+αS

p−1

, α ∈ [0, 1],

R(θ

p

+α

¯

h, θ

q

+β

¯

h) =

¸

(1 −α)R

pq

+βR

p−1,q−1

+(α −β)R

p−1,q

, α β,

(1 −β)R

pq

+αR

p−1,q−1

+(β −α)R

p,q−1

, α < β.

(2.24)

Thus, the LKF is completely determined by P

1

, Q

p

, S

p

, R

pq

, p, q = 0, 1, . . . , N.

The LKF condition V(x

t

) εx(t )

2

, ε > 0, is satisﬁed (Gu et al., 2003, p. 185) if S

p

> 0, p =

0, 1, . . . , N, and

¸

P

1

˜

Q

∗

˜

R +

˜

S

> 0, (2.25)

where

˜

Q = [Q

0

Q

1

· · · Q

N

],

˜

S = diag

¸

1

¯

hS

0

,

1

¯

hS

1

, . . . ,

1

¯

hS

N

¸

,

˜

R =

¸

¸

¸

¸

¸

¸

R

00

R

01

· · · R

0N

R

10

R

11

· · · R

1N

.

.

.

.

.

.

.

.

.

.

.

.

R

N0

R

N1

· · · R

NN

. (2.26)

To derive the LKF derivative condition, we note that

˙

S(ξ) =

1

¯

h(S

p−1

− S

p

)

,

˙

Q(ξ) =

1

¯

h(Q

p−1

− Q

p

)

,

∂

∂ξ

R(ξ, θ) +

∂

∂θ

R(ξ, θ) =

1

¯

h(R

p−1,q−1

− R

pq

)

. (2.27)

We have

W = ζ

v

Ξ

v

ζ

v

−

1

0

φ

(α)S

d

φ(α)dα −

1

0

¸

1

0

φ

(α)R

d

φ(β)dα

dβ

+2ζ

1

0

[D

s

+(1 −2α)D

a

]φ(α)

¯

h dα, (2.28)

where

ζ

v

= [x

(t ) ˙ x

(t ) x

(t −h) v

(t )],

10 of 15 E. FRIDMAN

Ξ

v

=

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

Ξ µ

¸

¸

¸

P

2

A

1

P

3

A

1

0

+µ

¸

Y

a

0

¸

¸

¸

¸

P

2

F

P

3

F

0

¸

−Y

a

0

¸

¸

¸

¸

P

2

H

P

3

H

0

∗ −µR

a

0 0 0

∗ ∗ −(1 − f )U 0 0

∗ ∗ ∗ −(1 −d)S

a

0

∗ ∗ ∗ ∗ −ρI

n

+ρ

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

E

0

0

n×n

E

1

µE

1

E

2

0

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

E

0

0

n×n

E

1

µE

1

E

2

0

, (2.29)

Ξ =

¸

¸

¸

Ψ

d

¸

P

2

A

1

P

3

A

1

−

¸

Q

N

0

¸

+Y

a

∗ −S

N

, P =

¸

P

1

0

P

2

P

3

¸

, Y

a

=

¸

P

2

¯

C

P

3

C

,

Ψ

d

= P

¸

0 I

A

0

−I

¸

+

¸

0 A

0

0 −I

¸

P +

¸

Q

0

+ Q

0

+ S

0

+ S

a

0

0 U +µR

a

¸

, (2.30)

φ

(α) = [x

(t −

¯

h +α

¯

h) x

(t −2

¯

h +α

¯

h) · · · x

(t − N

¯

h +α

¯

h)],

S

d

= diag{S

0

− S

1

, S

1

− S

2

, . . . , S

N−1

− S

N

},

R

d

=

¸

¸

¸

¸

¸

¸

R

d11

R

d12

· · · R

d1N

R

d21

R

d22

· · · R

d2N

.

.

.

.

.

.

.

.

.

.

.

.

R

dN1

R

dN2

· · · R

dNN

, R

dpq

=

¯

h(R

p−1,q−1

− R

pq

),

D

s

= [D

s

1

D

s

2

· · · D

s

N

], D

a

= [D

a

1

D

a

2

· · · D

a

N

],

D

s

p

=

¸

¸

¸

¯

h/2(R

0, p−1

+ R

0p

) −(Q

p−1

− Q

p

)

¯

h/2(Q

p−1

+ Q

p

)

−

¯

h/2(R

N, p−1

+ R

Np

)

and

D

a

p

=

¸

¸

¸

−

¯

h/2(R

0, p−1

− R

0p

)

−

¯

h/2(Q

p−1

− Q

p

)

¯

h/2(R

N, p−1

− R

Np

)

. (2.31a–i)

ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 11 of 15

Applying Proposition 5.21 of Gu et al. (2003) to (2.28) and Schur complements to the last term of

Ξ

v

, we conclude that W < 0 if the following LMI holds:

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

Ξ ψ

¸

¸

¸

P

2

F

P

3

F

0

¸

−Y

a

0

¸

¸

¸

¸

P

2

H

P

3

H

0

ρ

¸

¸

¸

E

0

0

E

1

D

s

D

a

∗ −µR

a

0 0 0 ρµE

1

0 0

∗ ∗ −(1 − f )U 0 0 ρE

2

0 0

∗ ∗ ∗ −(1 −d)S

a

0 0 0 0

∗ ∗ ∗ ∗ −ρI

n

0 0 0

∗ ∗ ∗ ∗ ∗ −ρI

n

0 0

∗ ∗ ∗ ∗ ∗ ∗ −R

d

− S

d

0

∗ ∗ ∗ ∗ ∗ ∗ ∗ −3S

d

< 0,

(2.32)

where

ψ = µ

¸

¸

¸

P

2

A

1

P

3

A

1

0

+µ

¸

Y

a

0

¸

. (2.33)

We thus obtain the following.

THEOREM 2 System (2.1) is asymptotically stable for all delays satisfying (2.3), if there exist

n×n matrices 0 < P

1

, P

2

, P

3

, R

a

, S

a

, Y

1a

, Y

2a

, US

p

= S

p

, Q

p

, R

pq

= R

qp

, p = 0, 1, . . . , N, q =

0, 1, . . . , N, and a scalar ρ > 0 such that LMIs (2.25) and (2.32) are satisﬁed with Y

a

= [Y

1a

Y

2a

]

and with the notation deﬁned in (2.30), (2.26), (2.31b–i) and (2.33).

REMARK 3 Similar to Fridman & Shaked (2006), it can be shown that the time-domain conditions

(2.32) are sufﬁcient for the frequency-domain conditions (2.11), where R

a

= X

1

X

1

, U = X

2

X

2

,

S

a

= X

3

X

3

and ρ = r

2

. Since Proposition 1 gives necessary conditions for feasibility of (2.11), while

(2.11) is necessary for feasibility of (2.32), the conditions of Proposition 1 are necessary for feasibility

of (2.32). We note that the conditions of Proposition 1 follow immediately from the feasibility of (2.32).

If LMI (2.32) is feasible, then the following LMI

¸

¸

¸

−P

3

− P

3

+U +µR

a

P

3

F µP

3

(A

1

+C)

∗ −(1 − f )U 0

∗ ∗ −µR

a

< 0,

C = P

−

3

Y

2a

(2.34)

holds. LMI (2.32) guarantees the stability of the difference equations (2.13) and (2.15) with constant

delays (Fridman, 2002).

REMARK 4 Since the LMIs (2.25) and (2.32) are afﬁne in the system matrices, the results of Theorem 2

can be applied to systems with polytopic-type uncertainties by solving LMIs in the polytope vertices

(Boyd et al., 1994).

12 of 15 E. FRIDMAN

REMARK 5 Stability conditions in the case of ˙ τ(t ) 1 follow from (2.25) and (2.32) by choosing

Y

a

= 0 and S

a

→ 0. In the case of fast-varying delays τ(t ) (without any constraints on the derivative

of τ(t )), the stability conditions have the form of (2.25) and (2.32), where the term µR

a

in Ξ should be

multiplied by 2 and where Y

a

= 0 and S

a

→ 0.

REMARK 6 Following De Oliveira & Skelton (2001), He et al. (2004) and Suplin et al. (2004), we can

introduce additional degrees of freedom by changing [x

(t )P

2

˙ x

(t )P

3

] in the right side of (2.21)

by the full-order vector ζ

v

, multiplied by the corresponding weighting matrices. This means that the

right side of the following equation can be added to

˙

V(x

t

) (additionally to (2.21)):

0 = 2[x

(t −h)P

4

v

1

(t )P

5

v

2

(t )P

6

v

3

(t )P

7

v

4

(t )P

8

]

×

¸

¸

¸

¸

¸

¸

¸

A

0

x(t ) − ˙ x(t ) +(A

1

+C

4

)x(t −h) +µ(A

1

+C

4

)v

1

(t ) + Fv

2

(t ) −C

4

v

3

(t ) + Hv

4

(t )

A

0

x(t ) − ˙ x(t ) +(A

1

+C

5

)x(t −h) +µ(A

1

+C

5

)v

1

(t ) + Fv

2

(t ) −C

5

v

3

(t ) + Hv

4

(t )

A

0

x(t ) − ˙ x(t ) +(A

1

+C

6

)x(t −h) +µ(A

1

+C

6

)v

1

(t ) + Fv

2

(t ) −C

6

v

3

(t ) + Hv

4

(t )

A

0

x(t ) − ˙ x(t ) +(A

1

+C

7

)x(t −h) +µ(A

1

+C

7

)v

1

(t ) + Fv

2

(t ) −C

7

v

3

(t ) + Hv

4

(t )

A

0

x(t ) − ˙ x(t ) +(A

1

+C

8

)x(t −h) +µ(A

1

+C

8

)v

1

(t ) + Fv

2

(t ) −C

8

v

3

(t ) + Hv

4

(t )

.

(2.35)

The LMI (2.32) in this case will take the form

Φ

0

+Φ

1

+Φ

1

< 0, (2.36)

where S

N

> 0, U > 0, R

a

> 0, S

a

> 0 and Φ

0

is the matrix in the left side of (2.32), while

Φ

1

=

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

¸

0

2n×2n

0

2n×n

0

2n×n

0

2n×n

0

2n×n

0

2n×n

0

P

4

[A

0

− I ] P

4

A

1

+Y

4

µP

4

A

1

+µY

4

P

4

F −Y

4

P

4

H 0

P

5

[A

0

− I ] P

5

A

1

+Y

5

µP

5

A

1

+µY

5

P

5

F −Y

5

P

5

H 0

P

6

[A

0

− I ] P

6

A

1

+Y

6

µP

6

A

1

+µY

6

P

6

F −Y

6

P

6

H 0

P

7

[A

0

− I ] P

7

A

1

+Y

7

µP

7

A

1

+µY

7

P

7

F −Y

7

P

7

H 0

P

8

[A

0

− I ] P

8

A

1

+Y

8

µP

8

A

1

+µY

8

P

8

F −Y

8

P

8

H 0

0 0 0 0 0 0 0

,

Y

i

= P

i

C

i

, i = 4, . . . , 8.

The above weighting matrices may lead to some improvement.

REMARK 7 During the last few years, new techniques for adding free weighting matrices have been

developed (see, e.g. He et al., 2005; Parlakci, 2006 and the references therein). All these work consider

simple Lyapunov functionals. The extension of these methods to the discretized Lyapunov functional

method may be the topic of future research.

The LMI conditions via discretized Lyapunov functional method are known to be numerically com-

plex, depending on the large number of decision variables (Gu, 1997; Gu et al., 1997; Han, 2005). The

descriptor discretized Lyapunov functional method (Fridman, 2006) adds to the existing method two

more n ×n decision variables P

2

and P

3

, that allow, however, to solve the design problems. Simpliﬁca-

tion of the discretized Lyapunov functional method is another direction for future research.

ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 13 of 15

EXAMPLE 1 Consider (2.1) with

A

0

=

¸

−2 0

0 −0.9

¸

, A

1

=

¸

−1 0

−1 −1

¸

, F =

¸

0.1 1

0 0.1

¸

,

H =

¸

0.2 0

0 0.2

¸

, E

0

= E

1

=

¸

1 0

0 1

¸

, E

2

=

¸

0.5 0

0 0.5

¸

. (2.37)

This system with F = diag{0.1, 0.1} and E

2

= 0 was analysed in Han (2005) by the discretized

Lyapunov functional method, where the following upper bound on the constant delay was found for

N = 3: 0 g = τ 2.12. Theorem 2 with N = 3 leads to a less restrictive bound for constant delays:

τ 2.4 and all g.

Consider now the case of H and E

2

given by (2.37) and time-varying g(t ) and τ(t ). Applying

the Euclidean matrix norm, we ﬁnd that F + HE

2

= 1.1 > 1 and thus the existing (direct)

Lyapunov methods cannot be applied. By Theorem 2 for N = 3, we ﬁnd that for all g(t ) the system is

asymptotically stable for τ(t ) from the following intervals:

f = d = 0, µ = 0, h 0.48, 0 τ 0.48,

f = d = 0.1, h = µ = 0.20, 0 τ(t ) 0.40,

f = d = 0.5, h = µ = 0.08, 0 τ(t ) 0.16.

Note that in this example, the free weighting matrices P

4

, . . . , P

8

, Y

4

, . . . , Y

8

of (2.36) do not

improve the results.

EXAMPLE 2 (Michiels & Vyhlidal, 2005) Consider

˙ x(t ) = Ax(t ) + BKx(t −τ(t )) + F

1

˙ x(t − g

1

(t )) + F

2

˙ x(t − g

2

(t )), (2.38)

where

F

1

=

¸

¸

¸

0 0.2 −0.4

−0.5 0.3 0

0.2 0.7 0

, F

2

=

¸

¸

¸

−0.3 −0.1 0

0 0.2 0

0.1 0 0.4

,

A =

¸

¸

¸

−4.8 4.7 3

0.1 1.4 −0.4

0.7 3.1 −1.5

, B =

¸

¸

¸

0.3

0.7

0.1

.

It was shown in Michiels & Vyhlidal (2005) that there exists K that stabilizes (2.38) with constant

delays τ ≡ 0.5, g

1

≡ 0.7 and g

2

≡ 1.7. Note that in this example F

1

+ F

2

= 1.22 > 1 and thus

the existing (direct) Lyapunov methods cannot be applied to the case of ‘time-varying’ g

1

or g

2

.

Consider now constant g

2

and time-varying g

1

(t ) and τ(t ) with ˙ g

1

f, ˙ τ d. We choose K =

[−0.3626 −6.7792 1.3247] (this gain was found by using the stabilization via descriptor discretized

Lyapunov functional, Fridman, 2006) and we analyse the asymptotic stability of the resulting closed-

loop system. By applying the extension of Theorem 2 and (2.36) with N = 3 to multiple delays g

1

and

g

2

, we obtain for all g

1

(t ) and g

2

the following stability intervals for τ(t ):

f = d = 0, h 0.12, µ = 0, 0 τ 0.12,

f = d = 0.1, h = 0.07, µ = 0.03, 0.04 τ(t ) 0.1.

14 of 15 E. FRIDMAN

In this example, the weighting matrices P

4

, . . . , P

8

, Y

4

, . . . , Y

8

improve the results; however, veriﬁca-

tion of the latter condition takes more computation time. Thus, without these matrices for f = d = 0

we ﬁnd 0 τ 0.04.

3. Conclusions

The input–output approach is extended to the stability analysis of linear neutral type systems with un-

certain time-varying delays and either norm-bounded or polytopic-type uncertainties. This allows a

restrictive assumption to be avoided on the sum of the norms of the matrices in the neutral part to be

less than 1. New sufﬁcient and necessary stability criteria are derived in the frequency and in the time

domains. These conditions are retarded-delay-dependent/neutral-delay-independent. The time-domain

criterion is based on the descriptor discretized Lyapunov functionals, which is known to be efﬁcient for

the solution of design problems. The method can be extended to L

2

-gain analysis.

The approach presented allows robust control theory to be developed to a wide class of neutral

uncertain systems with multiple time-varying delays. It gives insight to further development of the direct

Lyapunov method for neutral and more general descriptor systems with time-varying delays.

Funding

Kamea fund of Israel.

REFERENCES

BOYD, S., GHAOUI, L., FERON, E. & BALAKRISHNAN, V. (1994) Linear matrix inequalities in system and

control theory. Studies in Applied Mathematics, vol. 15. Philadelphia, PA: SIAM, pp. 1–205.

DE OLIVEIRA, M. C. & SKELTON, R. E. (2001) Stability test for constrained linear systems. Perspectives in

Robust Control (S. O. Reza Moheimani ed.). Lecture Notes in Control and Information Sciences, vol. 268.

London: Springer.

EL’SGOL’TS, L. & NORKIN, S. (1973) Introduction to the theory and applications of differential equations with

deviating arguments. Mathematics in Science and Engineering, vol. 105. New York: Academic Press.

FRIDMAN, E. (2001) NewLyapunov-Krasovskii functionals for stability of linear retarded and neutral type systems.

Syst. Control Lett., 43, 309–319.

FRIDMAN, E. (2002) Stability of linear descriptor systems with delay: a Lyapunov-based approach. J. Math. Anal.

Appl., 273, 24–44.

FRIDMAN, E. (2006) Descriptor discretized Lyapunov functional method: analysis and design. IEEE Trans. Autom.

Control, 51, 890–897.

FRIDMAN, E. & SHAKED, U. (2006) Input-output approach to stability and L

2

-gain analysis of systems with

time-varying delays. Syst. Control Lett., 55, 1041–1053.

GU, K. (1997) Discretized LMI set in the stability problem of linear time delay systems. Int. J. Control, 68,

923–934.

GU, K., KHARITONOV, V. & CHEN, J. (2003) Stability of Time-Delay Systems. Boston, MA: Birkhauser.

GU, K. & NICULESCU, S. (2001) Further remarks on additional dynamics in various model transformations of

linear delay systems. IEEE Trans. Autom. Control, 46, 497–500.

HALE, J. K. & LUNEL, S. M. (1993) Introduction to Functional Differential Equations. New York: Springer.

HAN, Q.-L. (2005) On stability of linear systems with mixed time delays: a discretized Lyapunov functional ap-

proach. Automatica, 41, 1209–1218.

HE, Y., WANG, Q.-G., LIN, Q.-G. & WU, M. (2005) Augmented Lyapunov functional and delay-dependent

stability criteria for neutral systems. Int. J. Robust Nonlinear Control, 15, 923–933.

ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 15 of 15

HE, Y., WU, M., SHE, J. & LIU, G. (2004) Delay-dependent robust stability criteria for uncertain neutral systems

with mixed delays. Syst. Control Lett., 51, 57–65.

HUANG, Y.-P. & ZHOU, K. (2000) Robust stability of uncertain time-delay systems. IEEE Trans. Autom. Control,

45, 2169–2173.

KAO, C.-Y. & LINCOLN, B. (2004) Simple stability criteria for systems with time-varying delays. Automatica, 40,

1429–1434.

KHARITONOV, V. & MELCHOR-AGUILAR, D. (2002) On delay-dependent stability conditions for time-varying

systems. Syst. Control Lett., 46, 173–180.

KOLMANOVSKII, V. & MYSHKIS, A. (1999) Applied Theory of Functional Differential Equations. Boston: Kluwer.

LIEN, C. H. (2005) Stability and stabilization criteria for a class of uncertain neutral systems with time-varying

delays. J. Optim. Theory Appl., 124, 637–657.

MICHIELS, W. & VYHLIDAL, T. (2005) An eigenvalue based approach for the stabilization of linear time-delay

systems of neutral type. Automatica, 41, 991–998.

NICULESCU, S. I. (2001) Delay Effects on Stability: A Robust Control Approach. Lecture Notes in Control and

Information Sciences, vol. 269. London: Springer.

PARK, J. (2002) Stability criterion for neutral differential systems with mixed multiple time-varying delay argu-

ments. Math. Comput. Simul., 59, 401–412.

PARLAKCI, M. (2006) Robust stability of uncertain time-varying state-delayed systems. IEE Proc. Control Theory

Appl., 153, 469–477.

SUPLIN, V., FRIDMAN, E. & SHAKED, U. (2004) H

∞

control of linear uncertain time-delay systems—a projection

approach. Proceedings of IEEE CDC 2004. Atlantis, Bahamas.

ZHANG, J., KNOPSE, C. & TSIOTRAS, P. (2001) Stability of time-delay systems: equivalence between Lyapunov

and small-gain conditions. IEEE Trans. Autom. Control, 46, 482–486.

means that P is symmetric and positive deﬁnite. for P ∈ Rn×n . (2. |η(t)| µ h. 2. are constant matrices.2) The uncertain delays τ (t) and g(t) are differentiable functions of the form τ (t) = h + η(t). Retarded-delay-dependent/neutral-delay-independent stability conditions 2. ∞) → C n with the norm v L 2 = 0 v(t) 2 dt . Rn denotes the n-dimensional Euclidean space with vector norm · .1 Problem formulation Consider a linear system x(t) = (A0 + H E 0 )x(t) + (A1 + H E 1 )x(t − τ (t)) + (F + H E 2 )x(t − g(t)). d and f . For a transfer function matrix of a stable system G(s). 0]). A1 . For simplicity. The latter method combines the discretized Lyapunov functional method of Gu (1997) with the descriptor model transformation (Fridman. The descriptor discretized Lyapunov functional method is chosen because of the power of the discretized method (which.2 of 15 E. To illustrate the efﬁciency of the new method for neutral systems. Also some ‘necessary conditions’ for the feasibility of the resulting stability conditions are deduced. we extend the method of Fridman (2006) to the case of slowly varying delays in the state and in the state derivative (where the delay derivative is bounded from above by a constant.1) (t) is a time-varying uncertain (2. The presented results may be easily ‘generalized to the case of any ﬁnite number of delays g(t)’. τ (t) ˙ d < 1. The symmetric elements of the symmetric matrix will be denoted by ∗. L 2 is 1/2 ∞ the space of square-integrable functions v: [0. The neutral delay g(t) does not have bound other than the derivative bound. FRIDMAN The presented method extends the input–output approach to uncertain neutral systems with time-varying delays and it leads to sufﬁcient frequency-domain stability conditions. F and E i . the time-domain results are derived via the descriptor discretized Lyapunov functional method (Fridman. i = −1. s ∈ C. Notation: Throughout the paper. n × n matrix that satisﬁes (t) (t) In . . We also denote xt (θ) = x(t + θ)(θ ∈ [−h − µ. ˙ ˙ where x(t) ∈ Rn . A denotes the Euclidean norm of an n × n (real or complex) matrix A. Rn×m is the set of all n × m real matrices and the notation P > 0. which is equal to the maximum singular value of A. We note that in Fridman (2006). differently to the simple Lyapunov functional methods. (2. the case of constant neutral delays and fast-varying state delays (without any constraints on the delay derivative) was studied.3) with the known bounds µ. g(t) ˙ f < 1. G ∞= −∞<w<∞ σ (B) is the spectral radius of matrix B (i. the superscript ‘ ’ stands for matrix transposition. 2001) and it efﬁciently solves design problems (for the ﬁrst time for the discretized method). 2. A0 . If τ (t) = g(t). we consider a single delay g(t). 2006). √ sup G(iw) .e. which is less than 1). 1. can analyse systems which are not stable without delays). then the stability criterion may be applied with f = d. i = 0. In the present paper. the maximum absolute value of its eigenvalues).

1) (denoted by f (t. we ﬁnd time-domain conditions by applying descriptor discretized Lyapunov functional method. y3 (t) = √ ˙ x(t). p. xt . We ﬁrst derive frequency-domain stability conditions and then deduce their implications. we consider the following forward system: x(t) = A0 x(t) + A1 x(t − h) + µ A1 v1 (t) + Fv2 (t) + H v4 (t).4) where · is any matrix norm. (2. v4 (t) = . xt ) − f (t. 2003 and the references therein). ˙ ¯ ˙ In some papers (see. where the condition (2. In the robust stability context..5a–d) with the feedback 1 v1 (t) = − µ v3 (t) = √ + s)ds. xt )) satisﬁes the Lipschitz condition with respect to xt with a ˙ ˙ constant less than 1 (see Kolmanovskii & Myshkis. 1− f 1−d y4 (t) = E 0 x(t) + E 1 x(t − h) + µE 1 v1 (t) + E 2 v2 (t). ˙ 1 1 y1 (t) = x(t). y2 (t) = √ ˙ x(t). robust stability of (2.ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 3 of 15 The stability analysis of systems with a neutral time-varying delay is a classical problem (El’sgol’ts & Norkin. where the parameter uncertainties and the variation of the time delay are taken into account. −h −h−η(t) y1 (t (2. this problem becomes especially important. 2005) F + H E 2 < 1. v2 (t) = √ 1 − f y2 (t − g(t)).1) with time-varying delay was studied under the restrictive assumption (Lien. Unfortunately. (2. Finally. where V > 0 is a Lyapunov functional. Park. xt . 2. 1999. In the present paper. it is directly concluded that the neutral system is asymptotically stable (without any references to the corresponding Lyapunov theorems). xt .4) guarantees the latter Lipschitz condition since ˙ ˙ ˙ ¯ ˙ ¯ f (t.4) is removed. The new method gives tools for solution of different robust control problems for a wide class of neutral systems. In the past. xt ) = (F + H E 2 )[x(t − g(t)) − x(t − g(t))] ( F + H ˙ E 2 ) x(t − g(t)) − x(t − g(t)) .2 Frequency-domain stability conditions Representing t−h x(t − τ (t)) = x(t − h) − t−h−η(t) x(s)ds ˙ and applying the input–output approach (see Gu et al. such a conclusion is not correct. 2002) from the fact that V < 0. The latter assumption (which becomes especially restrictive in the case of multiple neutral delays) allows the direct Lyapunov method to be applied where it is assumed that the right side of (2. we illustrate the efﬁciency of the proposed method by numerical examples.g. y4 (t). 336). e. 1973). Inequality (2.6) 1 − d y3 (t − τ (t)). we develop the input–output approach to the stability analysis of uncertain neutral type systems with time-varying delays. Further.

7a–d). . where x(t) appears in the ˙ derivative of the Lyapunov functional. while C = − A1 corresponds to τ -independent/τ -dependent result. 2001). where C. the input–output model (2. i = (2. y4 (t) = E 0 x(t) + E 1 x(t − h) + µE 1 v1 (t) + E 2 v2 (t). ˙ 1− f y3 (t) = √ 1 1−d (2. y4 ]. (2. 2001). .8) 0. (A1) The (parameterized) nominal system x(t) = A0 x(t) + (A1 + C)x(t − h) ˙ is asymptotically stable. ˙ where C is an arbitrary n × n matrix (which is equivalent to the parameterized model transformation Niculescu. ∀ t i = 1.1) with delay h 0 and non-negative η ∈ [0. (2004). (2. were taken to be zero. 3. we add to (2. . We represent the forward system in the following parameterized form: x(t) = A0 x(t) + (A1 + C)x(t − h) + µ( A1 + C)v1 (t) + Fv2 (t) ˙ − Cv3 (t) + H v4 (t). Let v = [v1 · · · v4 ] and y 1.4 of 15 E.6) ˙ has been introduced in Fridman & Shaked (2006). ˙ In the case of a retarded system with F = 0 and with τ 1.7a–d) can be written as y = Gv with transfer matrix G(s) = s In √ 1 s In 1− f √ 1 1−d In E 0 + E 1 e−hs F −C H] × (s I − A0 − (A1 + C)e−hs )−1 [µ( A1 + C) + 03n×n µE 1 03n×n E2 03n×2n 0n×2n . without any constraints on τ ). In the latter case. Assume that yi (t) = 0.5a) the left side of the equation C[x(t − h) + µv1 (t) − v3 (t)] = C x(t − h) − −h −h−η(t) x(t + s)ds − x(t − τ (t)) = 0. .7a–d). y1 and y2 correspond to the descriptor method (Fridman. µ].10) . vi and yi . FRIDMAN Similarly to He et al.6) and the results of the present paper are appropriate also to the system (2. The forward system (2. . . i = 2. We assume the following. C = 0 corresponds to systems with fast-varying delay τ (i. 4. (2. y1 (t) = x(t). The input–output model (2. Moreover. . ˙ ˙ 2006).7a–d) x(t). The following holds: vi L2 (2. 4. ˙ y2 (t) = √ 1 x(t). .9) = [y1 yi ··· L2 .e. Note that C = 0 corresponds to the moderately varying delay τ (t) with τ 1 (Fridman & Shaked.

1) is input–output stable. the system (2. since the nominal system is time invariant) if G ∞ < 1. R EMARK 1 In Section 2.15) . X 2 . θ ∈ [0.11) will be related to the decision variables of the LMIs (see Remark 3 below).3 Implications of the frequency-domain stability conditions Now. X 3 . 2003.e. Theorem 6. 2π ] .4 below sufﬁcient conditions for the feasibility of A1 and (2.1) is input–output stable (and thus asymptotically stable.13) (2.11) then (2. A stronger result may be obtained by scaling G.11) will be derived in terms of linear matrix inequalities (LMIs).14) 1 F x(t − g) = 0.. where η(t) and g(t) are differentiable functions. with constant delay g0 is stable. p. r In }G(s) diag{X 1 . The free matrices C. X 2 . 2.12) which is equivalent to the delay-independent stability of the following difference equations with constant delays g0 and g (Hale & Lunel. P ROPOSITION 1 If (2. Gu et al. −1 −1 −1 G X (s) = diag{X 1 . i. 286): x(t) − µ( A1 + C)x(t − g0 ) − √ 1 F x(t − g) = 0.3). i.11) holds.1) with delays given by (2. where σ0 = sup σ µ( A1 + C)eiθ0 + √ 1 F eiθ : θ0 . 1− f (2. 2003). 1 (ii) the eigenvalues of √1− f F are inside the unit circle. e. 2. we are in a position to formulate some necessary conditions for the feasibility of (2.g. r −1 In }.e. T HEOREM 1 Consider (2. and a scalar r = 0 such that GX ∞ < 1. then (i) the eigenvalues of µ( A1 + C) are inside the unit circle (µσ ( A1 + C) < 1).1. i = 1. X 3 . 1− f (2. If there exists an n × n matrix C such that A1 holds and there exist non-singular n × n matrices X i . the difference equation x(t) − µ( A1 + C)x(t − g0 ) = 0. 3.11). the difference equation x(t) − √ with constant delay g is stable. (iii) σ0 < 1. (2.ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 5 of 15 By the small-gain theorem (see. 1− f (2. X i and the scalar r of (2.

FRIDMAN Proof. e. where C is related to the decision variables of these LMIs (as. where additional dynamics of the transformed systems were analysed. 2003). may usually be recovered via an input–output approach (similar to Zhang et al. We will prove (iii) only. µk ].. i. in (2. ˙ . Necessary conditions for delay-dependent stability via different model transformations of linear retarded type systems with ‘small’ delays (where h = 0 and η ∈ [0. it follows that |λ| X 1a X √ 2a 1− f F eiθ a = λa. However.16) . X1 √X 2 1− f −1 [µ( A1 + C)X 1 −1 F X2 ] −∞<w<∞ sup iw X1 √X 2 1− f (iw In − A0 −(A1 +C)e−hiw )−1 −1 F X2 ] −1 × [µ( A1 + C)X 1 GX ∞ < 1..16). Then.e. √ F eiθ 1− f (2. The proof of (i) and (ii) is similar. derived via different direct Lyapunov methods.11) and for A1) are feasible if the difference equation (2. τk < 1. Assume that (2.34) below). µ]) were found in Gu & Niculescu (2001) and Kharitonov & Melchor-Aguilar (2002).17) Hence. 2001. the LMI conditions (which give sufﬁcient conditions for the frequency-domain condition (2. A simple condition of Kharitonov & Melchor-Aguilar (2002) m µk Ak < 1 for the system k=1 m x(t) = ˙ k=1 Ak x(t − τk (t)). A simple condition of Gu & Niculescu (2001) (for the constant delay case) µσ ( A1 ) < 1 coincides with (i) for the case of fast-varying delay. 2µ].g. |λ| < 1. for h = µ and τ ˙ 1 (i) guarantees the stability on the double interval τ (t) ∈ [0. R EMARK 2 We note that the existing LMI stability criteria. τk (t) ∈ [0.11) holds. Gu et al. (2. Let λ and a be an eigenvalue and an eigenvector of µ( A1 + C)eiθ0 + µ( A1 + C)eiθ0 + √ From (2. This completes the proof. In this case. 1− f = X1 √X 2 1− f λa = X1 √X 2 1− f −1 [µ( A1 + C)X 1 −1 F X2 ] X 1 eiθ0 a √ X 2 eiθ a 1− f < X 1 eiθ0 a √ X 2 eiθ a 1− f = X 1a √X 2 1− f a .15) is stable.6 of 15 E.

Since ∞ 0 ¯ v (t) X v(t)dt ∞ 0 ¯ y (t) X y(t)dt. The new descriptor discretized Lyapunov method of Fridman (2006) will be used. in the time domain we do not assume the asymptotic stability of the nominal system.19) for some n × n matrices Ra > 0.8).7a–d) W = d V (xt ) + (1 − f )y2 (t)U y2 (t) + µy1 (t)Ra y1 (t) dt +(1 − d)y3 (t)Sa y3 (t) + ρy4 (t)y4 (t) − (1 − f )v2 (t)U v2 (t) −µv1 (t)Ra v1 (t) − (1 − d)v3 (t)Sa v3 (t) − ρv4 (t)v4 (t) < −ε( x(t) 2 + x(t) ˙ 2 + v(t) 2 ).18) + −h x (t + ξ )S(ξ )x(t + ξ )dξ. the following condition along with (2. where the latter has the form µk Ak eiθk : θk ∈ [0. 2π ] < 1 sup σ k=1 and guarantees the stability on the double intervals τk ∈ [0.. (1 − d)Sa . 2. . ¯ X = diag{µRa . which are not stable without delays.1) (Gu et al.11) by using the complete Lyapunov–Krasovskii functional (LKF): 0 0 0 −h V (xt ) = x (t)P1 x(t) + 2x (t) 0 −h Q(ξ )x(t + ξ )dξ + P1 > 0. (1 − f )U.19) guarantees A1.ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 7 of 15 is more restrictive than (i) in the case of τk ˙ m 1. Sa and a scalar ρ > 0 guarantees the asymptotic stability of (2. 2µk ] for h k = µk . (2.4 Time-domain criterion for robust stability We will derive sufﬁcient conditions for (2. which for S ≡ 0 corresponds to necessary and sufﬁcient conditions for stability of the nominal system (2. Therefore. ρ In }. 2003). ξ )dsx(t + ξ )dξ (2. since this powerful method (differently from simple Lyapunov functionals) can be applied to the systems. ε > 0. Note that (2. 2003). which allows to solve the design problems. −h x (t + s)R(s. The time-domain results will be derived via the discretized Lyapunov functional method (Gu. U > 0.

θq−1 ]. Each small square is further divided into two triangles. where θ p = − p h. (2. we ﬁnd ˙ V (xt ) = ζ Ξ ζ + 2x (t) ˙ 0 0 −h 0 −h Q(ξ )x(t + ξ )dξ ∂ ∂ R(ξ. P3 x(t) ˙ ζ = x(t) . θ p−1 ] × [θq . of equal length h = h/N . 0] into N segments ¯ ¯ [θ p . x(t − h) Ψ Ξ = ∗ P= P1 P2 Ψ =P 0 A0 0 0 + −I 0 A0 −I P+ Q(0) + Q (0) + S(0) 0 . N . . P3 (A1 + C) −S(−h) 0 .22) µ( A1 + C)v1 (t) + Fv2 (t) − Cv3 (t) + H v4 (t) ¯ Q(−h) P2 (A1 + C) − 0 . . ξ )ds x(t + ξ )dξ + 2 ˙ −h x (t + ξ )S(ξ )x(t + ξ )dξ.21) where P2 and P3 are n × n matrices. FRIDMAN ˙ Differentiating V (xt ) along the trajectories of (2. 0] into N × N small squares [θ p .20). A0 x(t) − x(t) + (A1 + C)x(t − h) + µ( A1 + C)v1 (t) + Fv2 (t) − Cv3 (t) + H v4 (t) ˙ (2. which is equivalent to descriptor model transformation of Fridman (2001) and integrating by parts in (2. θ)x(t + θ )dθ − −h ˙ x (t + ξ ) S(ξ )x(t + ξ )dξ x (t)P3 ] ˙ ¯ ¯ µ( A1 + C)v1 (t) + Fv2 (t) − Cv3 (t) + H v4 (t) . θ) + R(ξ. . This divides the square [−h. .8 of 15 E. θ p−1 ]. 0] × [−h. we obtain that V is given by ˙ V (xt ) = 2x (t) P1 x(t) + ˙ 0 0 −h 0 −h 0 Q(ξ )x(t + ξ )dξ + 2x (t) 0 −h Q(ξ )x(t + ξ )dξ ˙ +2 0 = 2[x (t)P2 × ˙ Adding to V (x t ) the right side of the expression x (t)P3 ] ˙ −h x (t + s)R(s. 0 0 (2.7a–d). (2. ξ )]x(t + ξ )dξ 0 0 − 2x (t − h) + 2[x (t)P2 where −h R(−h.23a–c) We apply next the discretization of Gu (1997).20) ˙ ¯ ¯ ¯ A0 x(t) − x(t) + (A1 + C)x(t − h) + µ( A1 + C)v1 (t) + Fv2 (t) − Cv3 (t) + H v4 (t) ˙ . θ ) x(t + θ )dθ dξ ∂ξ ∂θ − −h x (t + ξ ) 0 + 2x (t) −h ˙ [− Q(ξ ) + R(0. p = 1. . Divide the delay interval [−h.

. . and P1 ∗ where ˜ Q = [Q 0 Q1 ··· Q N ]. ˜ S = diag R01 R11 . ¯ S(θ p + α h) = (1 − α)S p + αS p−1 . θ) + R(ξ. . ··· 1 1 1 . θ ) = . RN N 1 . S p . 1.. .. .q−1 . the LKF is completely determined by P1 . . p. is satisﬁed (Gu et al.. R pq . α < β. ¯ S0 h S1 ¯ ¯ SN h h R0N R1N . ˜ Q ˜+S R ˜ > 0. . α ∈ [0. q = 0. ¯ p−1 − Q p ) h(Q (2.. ε > 0.q . . . .27) . . (2. .24) (1 − α)R pq + β R p−1. Q p . p = 0.. .q−1 − R pq ) We have 1 1 1 W = ζ v Ξv ζ v − 1 0 φ (α)Sd φ(α)dα − φ (α)Rd φ(β)dα dβ 0 0 +2ζ where ζv = [x (t) 0 ¯ [D s + (1 − 2α)D a ]φ(α)h dα. RN 0 (2.q−1 + (α − β)R p−1. . 2003. . α β.26) To derive the LKF derivative condition. p.q−1 + (β − α)R p. ¯ ∂ξ ∂θ h(R p−1. RN 1 ··· ··· .28) x (t) ˙ x (t − h) v (t)].. θ) is chosen to be linear within each triangular: ¯ Q(θ p + α h) = (1 − α)Q p + α Q p−1 . 1.25) R00 R 10 ˜ R = . N . ¯ ¯ R(θ p + α h. (1 − β)R pq + α R p−1. . (2. (2. 185) if S p > 0. . N . ¯ p−1 − S p ) h(S ˙ Q(ξ ) = ∂ ∂ 1 R(ξ. . The LKF condition V (xt ) ε x(t) 2 . 1]. θq + β h) = Thus.ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 9 of 15 The continuous matrix functions Q(ξ ) and S(ξ ) are chosen to be linear within each segment and the continuous matrix function R(ξ. we note that ˙ S(ξ ) = 1 .

31a–i) . P3 C Q 0 + Q 0 + S0 + Sa 0 ··· 0 . P3 P2 H P3 H 0 0 0 0 −ρ In (2.30) ¯ ¯ φ (α) = [x (t − h + α h) ¯ ¯ x (t − 2h + α h) ¯ ¯ x (t − N h + α h)]. P+ P= P1 P2 0 . S1 − S2 .. p−1 + R N p ) D a ].29) Ψ Ξ = d ∗ Ψd = P Ya = ¯ P2 C . D a = [D1 D2 · · · N ¯ h/2(R0. . . U + µRa (2. . Rd N 1 Rd12 Rd22 . . s s a a D s = [D1 D2 · · · D s ]. Sd = diag{S0 − S1 . p−1 + R0 p ) − (Q p−1 − Q p ) ¯ Ds = h/2(Q p−1 + Q p ) and p ¯ −h/2(R N . . . N ¯ −h/2(R0. . Rd N N ¯ Rdpq = h(R p−1. p−1 − R0 p ) ¯ D a = −h/2(Q p−1 − Q p ) .q−1 − R pq ). Rd N 2 ··· ··· . p ¯ h/2(R N . p−1 − R N p ) (2. Rd11 R d21 Rd = . S N −1 − S N }. . . . .10 of 15 Ξ Ξv = ∗ ∗ ∗ ∗ P2 A1 Y µ P3 A1 + µ a 0 0 −µRa ∗ ∗ ∗ E. +ρ µE 1 µE 1 E2 E2 0 0 P2 A1 QN − 0 P3 A1 −S N 0 A0 I 0 + −I 0 A0 −I + Ya . ··· Rd1N Rd2N . FRIDMAN P2 F P3 F 0 0 −(1 − f )U ∗ ∗ −Ya 0 0 0 −(1 − d)Sa ∗ E0 E0 0n×n 0n×n E1 E1 . .

32) where P2 A1 Y ψ = µ P3 A1 + µ a 0 0 . . N . the conditions of Proposition 1 are necessary for feasibility of (2. T HEOREM 2 System (2. Sa . .32) guarantees the stability of the difference equations (2. we conclude that W < 0 if the following LMI holds: E0 P2 H P2 F −Ya Ds Da ψ Ξ P3 H ρ 0 P3 F 0 0 0 E1 ∗ −µRa 0 0 0 ρµE 1 0 0 ∗ ∗ −(1 − f )U 0 0 ρ E2 0 0 < 0. Sa = X 3 X 3 and ρ = r 2 . the results of Theorem 2 can be applied to systems with polytopic-type uncertainties by solving LMIs in the polytope vertices (Boyd et al. (2.32) is feasible. R pq = Rqp . .25) and (2. If LMI (2.30)..32) are sufﬁcient for the frequency-domain conditions (2. Y2a . . 1.32).32) are satisﬁed with Ya = [Y1a Y2a ] and with the notation deﬁned in (2. then the following LMI P3 F µP3 ( A1 + C) −P3 − P3 + U + µRa ∗ −(1 − f )U 0 < 0.32).32) are afﬁne in the system matrices.15) with constant delays (Fridman. .21 of Gu et al. .11) is necessary for feasibility of (2. 2002). . q = 0.25) and (2. N . R EMARK 3 Similar to Fridman & Shaked (2006). P3 .33) We thus obtain the following. where Ra = X 1 X 1 . Y1a .13) and (2. P2 .28) and Schur complements to the last term of Ξv .3). . it can be shown that the time-domain conditions (2. (2. R EMARK 4 Since the LMIs (2. 1994). Ra .33).26). Since Proposition 1 gives necessary conditions for feasibility of (2. . (2. if there exist n ×n matrices 0 < P1 .ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 11 of 15 Applying Proposition 5. Q p .32). We note that the conditions of Proposition 1 follow immediately from the feasibility of (2. LMI (2. and a scalar ρ > 0 such that LMIs (2.34) ∗ ∗ −µRa − C = P3 Y2a holds.11). ∗ 0 0 0 0 ∗ ∗ −(1 − d)Sa ∗ ∗ ∗ ∗ −ρ In 0 0 0 ∗ 0 0 ∗ ∗ ∗ ∗ −ρ In ∗ ∗ ∗ ∗ ∗ ∗ −Rd − Sd 0 ∗ ∗ ∗ ∗ ∗ ∗ ∗ −3Sd (2.11). U S p = S p . while (2. p = 0.31b–i) and (2.1) is asymptotically stable for all delays satisfying (2. (2003) to (2. (2. 1. U = X 2 X 2 .

multiplied by the corresponding weighting matrices.32) by choosing Ya = 0 and Sa → 0.21) ˙ by the full-order vector ζv .g.. e.35) . where S N > 0. This means that the ˙ right side of the following equation can be added to V (xt ) (additionally to (2. The LMI conditions via discretized Lyapunov functional method are known to be numerically complex. He et al. to solve the design problems.25) and (2. Sa > 0 and Φ0 is the matrix in the left side of (2. new techniques for adding free weighting matrices have been developed (see. Ra > 0. 2005. R EMARK 7 During the last few years. A0 x(t) − x(t) + ( A1 + C7 )x(t − h) + µ( A1 + C7 )v1 (t) + Fv2 (t) − C7 v3 (t) + H v4 (t) ˙ A0 x(t) − x(t) + ( A1 + C8 )x(t − h) + µ( A1 + C8 )v1 (t) + Fv2 (t) − C8 v3 (t) + H v4 (t) ˙ The LMI (2. R EMARK 6 Following De Oliveira & Skelton (2001). Simpliﬁcation of the discretized Lyapunov functional method is another direction for future research. The above weighting matrices may lead to some improvement.36) (2. 2006) adds to the existing method two more n × n decision variables P2 and P3 . FRIDMAN R EMARK 5 Stability conditions in the case of τ (t) ˙ 1 follow from (2.32).12 of 15 E. All these work consider simple Lyapunov functionals. The extension of these methods to the discretized Lyapunov functional method may be the topic of future research.21)): 0 = 2[x (t − h)P4 v1 (t)P5 v2 (t)P6 v3 (t)P7 v4 (t)P8 ] ˙ A0 x(t) − x(t) + ( A1 + C4 )x(t − h) + µ( A1 + C4 )v1 (t) + Fv2 (t) − C4 v3 (t) + H v4 (t) A0 x(t) − x(t) + ( A1 + C5 )x(t − h) + µ( A1 + C5 )v1 (t) + Fv2 (t) − C5 v3 (t) + H v4 (t) ˙ ˙ × A0 x(t) − x(t) + ( A1 + C6 )x(t − h) + µ( A1 + C6 )v1 (t) + Fv2 (t) − C6 v3 (t) + H v4 (t) . (2. 2006 and the references therein). i = 4.32) in this case will take the form Φ0 + Φ1 + Φ1 < 0. where the term µRa in Ξ should be multiplied by 2 and where Ya = 0 and Sa → 0. .25) and (2.32). . . 2005). . however. that allow. we can introduce additional degrees of freedom by changing [x (t)P2 x (t)P3 ] in the right side of (2. (2004) and Suplin et al.. In the case of fast-varying delays τ (t) (without any constraints on the derivative of τ (t)). 1997. depending on the large number of decision variables (Gu. 1997. 8. Gu et al. while 02n×n 02n×n 02n×n 02n×n 02n×n 0 02n×2n P [A − I ] P A + Y µP4 A1 + µY4 P4 F −Y4 P4 H 0 1 4 0 4 4 P5 [A0 − I ] P5 A1 + Y5 µP5 A1 + µY5 P5 F −Y5 P5 H 0 Φ1 = P6 [A0 − I ] P6 A1 + Y6 µP6 A1 + µY6 P6 F −Y6 P6 H 0 . U > 0. The descriptor discretized Lyapunov functional method (Fridman. the stability conditions have the form of (2. P [A − I ] P A + Y µP7 A1 + µY7 P7 F −Y7 P7 H 0 7 0 1 7 7 P8 [A0 − I ] P8 A1 + Y8 µP8 A1 + µY8 P8 F −Y8 P8 H 0 0 0 0 0 0 0 0 Yi = Pi Ci . He et al. Han. (2004). Parlakci.

. P8 . Y4 . Note that in this example.5 0.ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 13 of 15 E XAMPLE 1 Consider (2. where the following upper bound on the constant delay was found for N = 3: 0 g = τ 2.4 0. Y8 of (2. τ (t) 0.2 0 . 0. h = µ = 0. F2 = 0 0 0. g1 ≡ 0. .1.20.36) do not improve the results.7 3 1.5.48.5 0 . 0 0 0 τ 0.2 0 0 .37) and time-varying g(t) and τ (t).7 −0.7792 1.9 A1 = −1 −1 0 .1.1 −1. 0 0. we ﬁnd that F + H E 2 = 1. we ﬁnd that for all g(t) the system is asymptotically stable for τ (t) from the following intervals: f = d = 0.2 −0. .1 > 1 and thus the existing (direct) Lyapunov methods cannot be applied.7 .3626 −6.1 3.5.16. . h = 0.3247] (this gain was found by using the stabilization via descriptor discretized Lyapunov functional.03. A = 0. −1 1 0 . 0. we obtain for all g1 (t) and g2 the following stability intervals for τ (t): f = d = 0. By Theorem 2 for N = 3.8 4.5 0.5 (2. the free weighting matrices P4 . .1 0.1. E XAMPLE 2 (Michiels & Vyhlidal. Consider now the case of H and E 2 given by (2.07.7.7 0 0. f = d = 0. 0.40.2 −4.22 > 1 and thus the existing (direct) Lyapunov methods cannot be applied to the case of ‘time-varying’ g1 or g2 . 0 τ 0.1 0.12.3 −0.08. h 0. Note that in this example F1 + F2 = 1.4 .37) 0 . τ ˙ d.48.3 B = 0. Applying the Euclidean matrix norm. 0 0.1 0 0. f = d = 0.1 1 . h = µ = 0.3 0 . Fridman.4 F1 = −0.7 and g2 ≡ 1. µ = 0.4 and all g. .04 τ (t) 0.38) with constant delays τ ≡ 0. Consider now constant g2 and time-varying g1 (t) and τ (t) with g1 ˙ f.1} and E 2 = 0 was analysed in Han (2005) by the discretized Lyapunov functional method.12.1 (2. 0 1 F= E2 = 0.36) with N = 3 to multiple delays g1 and g2 . We choose K = [−0.12. µ = 0.2 E0 = E1 = This system with F = diag{0. 2005) Consider ˙ ˙ x(t) = Ax(t) + B K x(t − τ (t)) + F1 x(t − g1 (t)) + F2 x(t − g2 (t)). 0.1. . 0. ˙ where 0 0. −0. By applying the extension of Theorem 2 and (2. µ = 0.4 −0. . h 0. . τ (t) 0. 2006) and we analyse the asymptotic stability of the resulting closedloop system. Theorem 2 with N = 3 leads to a less restrictive bound for constant delays: τ 2. f = d = 0.38) It was shown in Michiels & Vyhlidal (2005) that there exists K that stabilizes (2.1) with A0 = H= −2 0 0.

(2003) Stability of Time-Delay Systems. (1997) Discretized LMI set in the stability problem of linear time delay systems. Control Lett. Y. Anal. Studies in Applied Mathematics. K. (2005) Augmented Lyapunov functional and delay-dependent stability criteria for neutral systems. J.. . (1973) Introduction to the theory and applications of differential equations with deviating arguments. S. D E O LIVEIRA . Automatica. . E. & BALAKRISHNAN .. (2005) On stability of linear systems with mixed time delays: a discretized Lyapunov functional approach. 15. It gives insight to further development of the direct Lyapunov method for neutral and more general descriptor systems with time-varying delays. Appl. (1993) Introduction to Functional Differential Equations. U. & N ICULESCU .. G U . Thus. 3. Mathematics in Science and Engineering. Control Lett. Control. Robust Nonlinear Control. J. 51. K. G U . P8 . 105. 46. Syst. R. (2006) Descriptor discretized Lyapunov functional method: analysis and design. V. F RIDMAN . 923–934. This allows a restrictive assumption to be avoided on the sum of the norms of the matrices in the neutral part to be less than 1. F RIDMAN . (2001) Stability test for constrained linear systems. S. F RIDMAN . M. veriﬁcation of the latter condition takes more computation time. MA: Birkhauser. . K HARITONOV. . however. The method can be extended to L 2 -gain analysis. The time-domain criterion is based on the descriptor discretized Lyapunov functionals. FRIDMAN In this example. PA: SIAM. New sufﬁcient and necessary stability criteria are derived in the frequency and in the time domains. 15. M. New York: Springer. Q. & S KELTON . the weighting matrices P4 . (2001) New Lyapunov-Krasovskii functionals for stability of linear retarded and neutral type systems. 43.. . L IN . F RIDMAN . 273.. London: Springer. & L UNEL . Control. 1209–1218. pp. E. L.-L. E. (2001) Further remarks on additional dynamics in various model transformations of linear delay systems. IEEE Trans. H AN . (2006) Input-output approach to stability and L 2 -gain analysis of systems with time-varying delays. vol. 309–319. Reza Moheimani ed. & W U . E. without these matrices for f = d = 0 we ﬁnd 0 τ 0.. E. J. Syst. New York: Academic Press. 55. J. V.-G. (2002) Stability of linear descriptor systems with delay: a Lyapunov-based approach. Int. (1994) Linear matrix inequalities in system and control theory. Q. vol.-G. 1–205. M. vol. Autom. IEEE Trans. S. . 41. The approach presented allows robust control theory to be developed to a wide class of neutral uncertain systems with multiple time-varying delays. C. & C HEN . F ERON . Lecture Notes in Control and Information Sciences. Funding Kamea fund of Israel. E L’ SGOL’ TS . J. 497–500. . H ALE . These conditions are retarded-delay-dependent/neutral-delay-independent. & N ORKIN . Q. & S HAKED . 68. Perspectives in Robust Control (S. 890–897. G U . 923–933. . Autom. Control. WANG . which is known to be efﬁcient for the solution of design problems. Y8 improve the results. H E . S. K. L. Math. 24–44. Philadelphia. G HAOUI . K. Conclusions The input–output approach is extended to the stability analysis of linear neutral type systems with uncertain time-varying delays and either norm-bounded or polytopic-type uncertainties.). R EFERENCES B OYD . 268. 1041–1053. ..04. O. E..14 of 15 E. Int. Y4 . Boston.

Syst. & S HAKED . V. 469–477. M. 45. M. 482–486. (2000) Robust stability of uncertain time-delay systems. IEEE Trans. Control Lett.. & L INCOLN . Automatica. Y. Proceedings of IEEE CDC 2004. Boston: Kluwer. C. Syst. Comput. 637–657. J. H UANG . . J. 46. C. K. 51. 269. W. K AO . (2004) Simple stability criteria for systems with time-varying delays. Automatica. J. (2001) Delay Effects on Stability: A Robust Control Approach. Lecture Notes in Control and Information Sciences. K NOPSE ..-P. & L IU . (2006) Robust stability of uncertain time-varying state-delayed systems. T.. & M ELCHOR -AGUILAR .ROBUST STABILITY OF LINEAR NEUTRAL SYSTEMS 15 of 15 H E . 59. Control Lett. L IEN . 40. J. (2005) Stability and stabilization criteria for a class of uncertain neutral systems with time-varying delays.. (2004) Delay-dependent robust stability criteria for uncertain neutral systems with mixed delays. N ICULESCU .. 124. G. Y. 57–65. Control Theory Appl. U. 173–180. & Z HOU . IEEE Trans. Theory Appl.-Y. Control. Optim. 991–998. PARK . 153. Z HANG . Bahamas. M ICHIELS .. P. Atlantis. 2169–2173. Math. E. Control. S UPLIN . & M YSHKIS . IEE Proc. 1429–1434. I. H. 46.. S HE . (2005) An eigenvalue based approach for the stabilization of linear time-delay systems of neutral type. Autom. C. (2004) H∞ control of linear uncertain time-delay systems—a projection approach.. V. 41. B. A. (2002) On delay-dependent stability conditions for time-varying systems. S.. W U . Autom. Simul. (2001) Stability of time-delay systems: equivalence between Lyapunov and small-gain conditions. PARLAKCI . 401–412. & V YHLIDAL . vol. D. & T SIOTRAS . F RIDMAN . V. London: Springer. KOLMANOVSKII . K HARITONOV. (2002) Stability criterion for neutral differential systems with mixed multiple time-varying delay arguments. (1999) Applied Theory of Functional Differential Equations.

- lyapnv
- Time-Delay Systems an Overview of Some Recent Advances and Open Problems
- Adaptive Coordinated Control for Hot Strip Finishing Mills
- IJEART06520
- nonlinear control systems analysis and design
- Agarwal Priyanshu Digital Control Project Report
- the pole and zeros.pdf
- 6010
- lec5
- Sampled-Data Modeling and Analysis of PWM DC-DC Converters Part I Closed-Loop Circuits
- Cabeq_2010_03_za_web_03
- Assignment 1
- Stabilization of Nonlinear Time-Varying Systems%3a a Control Lyapunov Function Approach
- Lecture 11
- 34
- Paul C. Bressloff and Zachary P. Kilpatrick- Two-Dimensional Bumps in Piecewise Smooth Neural Fields with Synaptic Depression
- lcs 9
- Evidence combination from an evolutionary game theory perspective
- Sample
- Grimm - Linear LMI-Based External Anti-windup Augmentation for Stable Linear Systems
- Classical Control Revision[1]
- Cpc Report Chapter 1
- Mathematical Modeling and Nonlinear Controller Design for a Novel Electrohydraulic Power Steering System
- IJSCMC
- A Novel Data-driven Control Approach for a Class of discrete time non linear systems
- Zhiyong Xu, Yaroslav V. Kartashov and Lluis Torner- Upper threshold for stability of multipole-mode solitons in nonlocal nonlinear media
- International Journal of Information Technology, Control and Automation (IJITCA)
- IJSCMC
- [eng]advanced package training scaffolding 14.pdf
- Huang Multi Loop Control

Sign up to vote on this title

UsefulNot usefulClose Dialog## Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

Loading